transcribe-mlx

Local audio transcription to JSON using MLX Whisper. Runs on Apple Silicon.

Install

pip install transcribe-mlx

Usage

transcribe-mlx "audio.mp3" "output.json"
transcribe-mlx "audio.mp3" "output.json" --model mlx-community/whisper-large-v3-mlx

Output is saved as a JSON file with segments, words, and timestamps.

First run: the model (~1–3 GB) is downloaded automatically from HuggingFace and cached locally. Subsequent runs use the cache.

Arguments

Options

Requirements

• Python 3.11+
• Apple Silicon (MLX requires Metal)
• FFmpeg (brew install ffmpeg)

Development

git clone https://github.com/felipeinfantino/transcribe-mlx
cd transcribe-mlx
poetry install
poetry run transcribe-mlx "inputs/audio.mp3" "outputs/result.json"

Extending with a new package

poetry add [packagename]

# Check the CLI is still working
poetry run transcribe-mlx "inputs/audio.mp3" "outputs/result.json"

# Verify lockfile is clean
poetry lock

# Bump version
poetry version patch   # or minor / major

Git build versioning

Let x.y.z be the new bumped version:

git tag v[x.y.z]
git push origin main --tags

License

MIT